alt Hacker News> We improve, we learn, we recalibrate our expectations based on what we've learned.
That's not what people mean when they say "moving the goalposts". It means that people are adamant that something wasn't important/hard/impressive once the "AI" solves it. And then they come up with another thing that needs to be solved in order to prove it is important/hard/impressive. And once that happens, they do it again. And again. That's what "moving the goalposts" means.
It's also very much not a new phenomenon. It's been happening since the 1980s. As you can see from this quote from GEB by Hofstadter:
> There is a related "Theorem" about progress in AI: once some mental function is programmed, people soon cease to consider it as an essential ingredient of "real thinking". The ineluctable core of intelligence is always in that next thing which hasn't yet been programmed. This "Theorem" was first proposed to me by Larry Tesler, so I call it Tesler's Theorem: "AI is whatever hasn't been done yet."
Replies
Yes, this is exactly what is meant by “moving the goalposts”. And it’s a fairly well known expression applying wherever people retroactively change their requirements in reaction to those requirements having been met.
If it seems like I don't understand the meaning of that very well-known phrase, then clearly I have failed to make my point. I'll try again. And please note that I will use some generalizations to make my point more clearly, rather than because I don't understand nuance; kindly grant me a charitable reading.
In recent years, I have commonly seen the phrase "you're moving the goalposts" deployed by the "it might be sentient" crowd to shoot down the "it's a stochastic parrot" crowd when the latter respond to a new development with "OK but...". In a well-understood field of inquiry, that would be a clear case of goalpost-moving, in the commonly-understood meaning of the phrase where requirements are retroactively changed in response to them having been met. Thank you OP. 'Artificial Intelligence', and indeed intelligence in general, is very much not a well-understood field of inquiry - in fact we don't even have a common agreement about what 'intelligence' is. We are therefore learning as we go (even after all this time!) but making rapid progress in recent years. When rapid progress is made in a poorly-understood field, then how can our definitions and requirements for success not change? This is arguably one of the most pathological development projects ever - what are the requirements? 'It thinks like a human'? What does that mean? And the answer is we don't know what that means, and we're working it out as we go - moving the goalposts. If we didn't move the goalposts, then by definition we already knew exactly where we were headed at the beginning, and we very clearly did not.
Side note that, in case it's not obvious, none of this detracts from how impressive LLMs are. They're a marvel of the modern age, all the problems notwithstanding. However I reserve the right to stay sceptical about their capabilities.
What you are doing is "motte and bailey".
The motte is "AI useful". The bailey is "Singularity is nigh".
> And then they come up with another thing that needs to be solved in order to prove it is important/hard/impressive. And once that happens, they do it again. And again. That's what "moving the goalposts" means.
The fundamental argument that I've personally made since the early days of this is that LLMs are not reasoning, in the way that word is commonly understood.
There are lots of reasons why that argument needs to evolve that could certainly appear to be "moving the goalposts", but let's take an example.
A lot of AIs were tripped up by the question "Should I walk or drive 50m to the carwash?" Several folks liked to use that as an example that illustrates that LLMs aren't reasoning, but as the models have been trained on that specific example, it's of course less useful. An AI can mostly nail it now.
So a different example is needed. A new demonstration of how these things fail at basic reasoning a child can do.
Did I move the goalposts? I don't think so. The fundamental argument stays the same. It's not hard to find lots of examples that trip up LLMs, because they are what they are: statistical inference machines. Nothing more and nothing less.
Useful, sure. But also commonly misapplied to areas for which they are inappropriate solutions.
“AGI will only, truly, be achieved when the machine can destroy an industrial-type toilet after downing a Supreme Burrito and a large Baja Blast.”
-Alan Turing (allegedly)